{ "version": 0.1, "quantization": "fp8", "dtype": "bfloat16", "vocab_size": 128256, "rank": 0, "tensor_parallel": 1, "pipeline_parallel": 1, "vocab_embedding": { "weight": "model.embed_tokens.weight" }, "positional_embedding": null, "ln_embed": null, "layers": [ { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.0.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.0.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.0.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.0.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.0.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.0.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.0.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.0.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.0.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.0.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.0.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.0.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.0.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.0.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.0.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.0.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.0.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.0.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.0.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.0.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.0.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.0.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.0.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.1.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.1.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.1.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.1.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.1.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.1.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.1.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.1.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.1.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.1.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.1.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.1.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.1.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.1.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.1.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.1.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.1.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.1.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.1.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.1.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.1.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.1.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.1.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.2.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.2.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.2.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.2.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.2.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.2.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.2.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.2.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.2.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.2.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.2.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.2.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.2.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.2.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.2.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.2.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.2.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.2.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.2.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.2.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.2.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.2.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.2.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.3.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.3.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.3.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.3.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.3.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.3.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.3.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.3.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.3.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.3.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.3.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.3.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.3.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.3.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.3.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.3.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.3.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.3.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.3.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.3.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.3.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.3.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.3.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.4.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.4.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.4.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.4.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.4.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.4.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.4.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.4.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.4.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.4.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.4.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.4.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.4.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.4.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.4.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.4.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.4.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.4.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.4.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.4.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.4.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.4.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.4.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.5.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.5.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.5.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.5.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.5.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.5.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.5.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.5.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.5.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.5.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.5.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.5.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.5.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.5.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.5.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.5.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.5.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.5.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.5.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.5.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.5.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.5.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.5.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.6.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.6.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.6.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.6.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.6.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.6.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.6.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.6.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.6.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.6.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.6.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.6.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.6.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.6.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.6.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.6.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.6.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.6.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.6.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.6.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.6.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.6.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.6.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.7.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.7.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.7.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.7.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.7.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.7.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.7.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.7.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.7.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.7.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.7.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.7.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.7.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.7.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.7.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.7.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.7.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.7.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.7.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.7.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.7.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.7.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.7.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.8.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.8.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.8.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.8.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.8.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.8.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.8.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.8.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.8.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.8.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.8.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.8.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.8.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.8.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.8.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.8.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.8.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.8.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.8.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.8.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.8.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.8.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.8.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.9.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.9.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.9.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.9.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.9.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.9.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.9.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.9.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.9.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.9.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.9.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.9.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.9.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.9.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.9.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.9.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.9.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.9.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.9.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.9.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.9.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.9.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.9.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.10.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.10.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.10.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.10.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.10.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.10.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.10.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.10.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.10.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.10.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.10.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.10.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.10.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.10.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.10.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.10.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.10.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.10.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.10.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.10.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.10.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.10.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.10.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.11.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.11.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.11.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.11.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.11.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.11.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.11.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.11.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.11.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.11.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.11.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.11.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.11.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.11.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.11.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.11.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.11.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.11.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.11.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.11.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.11.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.11.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.11.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.12.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.12.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.12.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.12.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.12.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.12.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.12.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.12.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.12.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.12.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.12.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.12.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.12.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.12.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.12.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.12.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.12.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.12.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.12.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.12.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.12.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.12.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.12.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.13.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.13.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.13.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.13.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.13.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.13.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.13.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.13.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.13.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.13.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.13.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.13.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.13.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.13.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.13.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.13.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.13.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.13.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.13.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.13.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.13.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.13.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.13.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.14.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.14.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.14.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.14.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.14.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.14.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.14.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.14.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.14.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.14.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.14.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.14.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.14.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.14.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.14.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.14.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.14.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.14.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.14.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.14.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.14.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.14.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.14.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.15.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.15.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.15.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.15.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.15.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.15.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.15.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.15.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.15.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.15.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.15.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.15.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.15.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.15.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.15.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.15.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.15.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.15.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.15.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.15.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.15.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.15.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.15.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.16.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.16.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.16.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.16.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.16.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.16.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.16.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.16.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.16.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.16.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.16.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.16.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.16.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.16.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.16.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.16.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.16.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.16.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.16.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.16.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.16.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.16.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.16.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.17.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.17.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.17.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.17.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.17.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.17.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.17.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.17.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.17.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.17.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.17.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.17.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.17.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.17.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.17.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.17.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.17.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.17.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.17.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.17.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.17.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.17.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.17.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.18.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.18.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.18.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.18.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.18.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.18.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.18.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.18.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.18.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.18.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.18.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.18.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.18.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.18.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.18.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.18.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.18.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.18.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.18.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.18.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.18.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.18.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.18.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.19.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.19.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.19.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.19.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.19.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.19.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.19.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.19.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.19.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.19.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.19.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.19.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.19.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.19.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.19.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.19.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.19.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.19.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.19.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.19.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.19.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.19.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.19.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.20.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.20.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.20.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.20.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.20.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.20.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.20.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.20.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.20.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.20.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.20.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.20.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.20.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.20.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.20.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.20.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.20.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.20.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.20.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.20.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.20.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.20.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.20.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.21.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.21.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.21.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.21.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.21.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.21.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.21.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.21.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.21.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.21.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.21.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.21.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.21.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.21.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.21.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.21.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.21.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.21.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.21.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.21.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.21.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.21.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.21.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.22.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.22.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.22.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.22.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.22.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.22.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.22.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.22.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.22.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.22.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.22.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.22.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.22.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.22.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.22.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.22.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.22.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.22.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.22.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.22.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.22.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.22.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.22.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.23.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.23.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.23.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.23.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.23.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.23.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.23.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.23.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.23.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.23.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.23.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.23.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.23.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.23.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.23.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.23.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.23.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.23.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.23.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.23.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.23.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.23.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.23.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.24.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.24.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.24.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.24.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.24.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.24.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.24.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.24.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.24.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.24.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.24.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.24.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.24.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.24.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.24.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.24.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.24.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.24.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.24.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.24.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.24.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.24.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.24.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.25.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.25.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.25.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.25.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.25.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.25.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.25.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.25.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.25.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.25.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.25.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.25.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.25.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.25.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.25.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.25.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.25.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.25.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.25.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.25.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.25.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.25.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.25.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.26.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.26.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.26.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.26.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.26.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.26.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.26.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.26.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.26.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.26.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.26.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.26.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.26.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.26.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.26.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.26.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.26.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.26.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.26.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.26.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.26.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.26.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.26.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.27.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.27.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.27.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.27.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.27.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.27.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.27.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.27.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.27.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.27.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.27.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.27.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.27.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.27.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.27.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.27.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.27.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.27.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.27.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.27.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.27.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.27.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.27.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.28.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.28.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.28.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.28.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.28.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.28.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.28.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.28.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.28.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.28.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.28.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.28.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.28.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.28.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.28.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.28.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.28.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.28.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.28.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.28.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.28.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.28.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.28.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.29.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.29.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.29.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.29.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.29.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.29.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.29.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.29.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.29.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.29.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.29.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.29.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.29.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.29.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.29.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.29.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.29.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.29.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.29.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.29.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.29.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.29.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.29.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.30.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.30.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.30.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.30.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.30.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.30.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.30.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.30.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.30.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.30.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.30.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.30.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.30.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.30.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.30.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.30.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.30.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.30.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.30.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.30.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.30.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.30.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.30.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.31.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.31.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.31.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.31.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.31.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.31.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.31.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.31.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.31.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.31.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.31.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.31.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.31.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.31.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.31.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.31.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.31.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.31.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.31.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.31.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.31.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.31.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.31.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.32.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.32.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.32.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.32.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.32.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.32.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.32.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.32.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.32.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.32.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.32.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.32.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.32.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.32.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.32.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.32.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.32.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.32.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.32.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.32.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.32.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.32.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.32.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.33.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.33.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.33.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.33.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.33.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.33.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.33.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.33.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.33.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.33.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.33.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.33.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.33.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.33.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.33.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.33.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.33.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.33.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.33.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.33.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.33.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.33.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.33.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.34.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.34.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.34.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.34.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.34.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.34.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.34.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.34.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.34.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.34.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.34.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.34.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.34.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.34.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.34.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.34.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.34.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.34.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.34.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.34.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.34.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.34.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.34.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.35.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.35.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.35.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.35.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.35.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.35.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.35.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.35.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.35.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.35.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.35.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.35.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.35.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.35.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.35.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.35.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.35.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.35.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.35.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.35.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.35.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.35.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.35.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.36.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.36.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.36.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.36.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.36.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.36.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.36.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.36.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.36.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.36.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.36.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.36.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.36.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.36.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.36.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.36.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.36.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.36.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.36.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.36.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.36.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.36.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.36.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.37.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.37.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.37.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.37.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.37.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.37.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.37.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.37.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.37.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.37.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.37.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.37.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.37.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.37.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.37.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.37.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.37.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.37.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.37.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.37.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.37.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.37.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.37.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.38.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.38.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.38.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.38.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.38.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.38.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.38.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.38.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.38.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.38.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.38.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.38.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.38.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.38.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.38.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.38.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.38.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.38.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.38.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.38.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.38.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.38.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.38.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.39.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.39.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.39.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.39.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.39.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.39.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.39.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.39.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.39.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.39.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.39.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.39.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.39.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.39.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.39.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.39.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.39.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.39.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.39.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.39.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.39.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.39.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.39.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.40.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.40.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.40.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.40.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.40.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.40.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.40.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.40.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.40.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.40.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.40.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.40.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.40.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.40.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.40.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.40.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.40.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.40.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.40.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.40.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.40.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.40.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.40.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.41.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.41.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.41.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.41.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.41.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.41.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.41.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.41.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.41.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.41.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.41.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.41.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.41.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.41.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.41.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.41.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.41.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.41.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.41.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.41.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.41.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.41.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.41.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.42.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.42.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.42.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.42.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.42.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.42.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.42.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.42.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.42.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.42.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.42.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.42.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.42.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.42.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.42.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.42.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.42.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.42.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.42.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.42.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.42.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.42.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.42.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.43.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.43.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.43.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.43.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.43.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.43.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.43.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.43.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.43.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.43.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.43.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.43.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.43.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.43.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.43.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.43.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.43.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.43.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.43.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.43.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.43.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.43.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.43.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.44.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.44.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.44.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.44.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.44.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.44.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.44.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.44.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.44.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.44.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.44.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.44.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.44.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.44.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.44.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.44.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.44.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.44.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.44.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.44.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.44.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.44.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.44.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.45.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.45.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.45.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.45.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.45.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.45.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.45.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.45.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.45.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.45.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.45.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.45.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.45.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.45.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.45.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.45.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.45.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.45.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.45.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.45.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.45.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.45.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.45.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.46.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.46.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.46.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.46.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.46.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.46.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.46.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.46.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.46.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.46.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.46.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.46.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.46.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.46.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.46.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.46.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.46.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.46.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.46.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.46.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.46.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.46.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.46.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.47.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.47.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.47.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.47.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.47.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.47.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.47.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.47.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.47.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.47.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.47.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.47.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.47.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.47.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.47.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.47.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.47.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.47.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.47.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.47.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.47.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.47.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.47.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.48.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.48.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.48.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.48.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.48.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.48.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.48.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.48.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.48.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.48.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.48.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.48.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.48.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.48.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.48.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.48.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.48.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.48.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.48.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.48.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.48.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.48.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.48.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.49.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.49.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.49.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.49.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.49.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.49.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.49.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.49.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.49.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.49.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.49.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.49.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.49.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.49.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.49.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.49.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.49.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.49.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.49.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.49.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.49.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.49.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.49.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.50.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.50.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.50.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.50.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.50.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.50.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.50.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.50.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.50.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.50.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.50.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.50.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.50.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.50.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.50.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.50.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.50.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.50.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.50.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.50.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.50.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.50.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.50.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.51.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.51.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.51.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.51.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.51.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.51.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.51.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.51.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.51.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.51.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.51.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.51.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.51.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.51.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.51.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.51.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.51.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.51.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.51.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.51.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.51.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.51.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.51.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.52.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.52.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.52.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.52.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.52.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.52.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.52.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.52.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.52.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.52.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.52.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.52.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.52.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.52.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.52.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.52.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.52.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.52.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.52.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.52.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.52.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.52.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.52.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.53.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.53.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.53.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.53.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.53.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.53.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.53.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.53.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.53.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.53.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.53.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.53.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.53.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.53.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.53.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.53.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.53.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.53.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.53.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.53.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.53.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.53.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.53.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.54.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.54.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.54.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.54.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.54.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.54.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.54.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.54.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.54.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.54.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.54.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.54.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.54.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.54.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.54.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.54.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.54.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.54.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.54.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.54.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.54.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.54.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.54.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.55.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.55.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.55.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.55.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.55.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.55.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.55.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.55.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.55.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.55.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.55.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.55.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.55.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.55.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.55.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.55.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.55.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.55.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.55.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.55.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.55.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.55.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.55.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.56.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.56.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.56.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.56.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.56.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.56.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.56.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.56.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.56.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.56.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.56.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.56.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.56.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.56.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.56.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.56.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.56.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.56.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.56.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.56.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.56.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.56.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.56.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.57.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.57.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.57.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.57.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.57.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.57.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.57.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.57.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.57.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.57.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.57.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.57.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.57.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.57.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.57.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.57.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.57.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.57.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.57.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.57.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.57.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.57.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.57.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.58.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.58.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.58.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.58.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.58.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.58.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.58.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.58.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.58.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.58.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.58.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.58.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.58.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.58.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.58.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.58.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.58.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.58.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.58.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.58.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.58.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.58.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.58.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.59.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.59.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.59.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.59.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.59.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.59.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.59.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.59.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.59.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.59.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.59.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.59.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.59.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.59.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.59.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.59.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.59.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.59.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.59.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.59.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.59.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.59.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.59.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.60.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.60.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.60.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.60.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.60.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.60.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.60.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.60.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.60.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.60.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.60.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.60.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.60.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.60.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.60.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.60.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.60.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.60.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.60.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.60.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.60.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.60.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.60.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.61.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.61.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.61.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.61.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.61.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.61.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.61.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.61.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.61.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.61.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.61.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.61.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.61.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.61.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.61.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.61.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.61.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.61.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.61.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.61.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.61.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.61.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.61.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.62.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.62.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.62.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.62.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.62.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.62.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.62.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.62.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.62.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.62.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.62.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.62.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.62.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.62.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.62.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.62.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.62.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.62.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.62.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.62.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.62.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.62.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.62.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.63.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.63.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.63.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.63.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.63.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.63.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.63.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.63.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.63.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.63.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.63.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.63.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.63.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.63.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.63.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.63.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.63.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.63.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.63.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.63.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.63.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.63.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.63.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.64.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.64.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.64.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.64.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.64.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.64.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.64.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.64.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.64.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.64.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.64.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.64.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.64.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.64.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.64.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.64.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.64.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.64.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.64.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.64.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.64.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.64.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.64.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.65.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.65.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.65.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.65.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.65.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.65.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.65.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.65.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.65.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.65.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.65.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.65.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.65.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.65.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.65.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.65.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.65.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.65.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.65.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.65.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.65.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.65.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.65.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.66.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.66.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.66.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.66.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.66.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.66.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.66.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.66.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.66.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.66.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.66.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.66.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.66.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.66.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.66.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.66.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.66.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.66.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.66.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.66.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.66.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.66.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.66.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.67.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.67.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.67.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.67.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.67.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.67.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.67.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.67.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.67.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.67.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.67.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.67.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.67.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.67.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.67.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.67.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.67.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.67.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.67.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.67.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.67.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.67.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.67.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.68.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.68.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.68.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.68.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.68.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.68.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.68.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.68.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.68.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.68.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.68.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.68.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.68.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.68.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.68.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.68.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.68.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.68.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.68.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.68.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.68.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.68.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.68.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.69.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.69.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.69.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.69.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.69.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.69.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.69.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.69.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.69.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.69.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.69.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.69.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.69.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.69.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.69.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.69.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.69.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.69.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.69.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.69.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.69.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.69.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.69.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.70.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.70.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.70.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.70.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.70.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.70.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.70.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.70.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.70.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.70.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.70.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.70.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.70.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.70.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.70.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.70.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.70.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.70.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.70.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.70.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.70.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.70.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.70.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.71.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.71.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.71.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.71.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.71.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.71.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.71.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.71.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.71.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.71.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.71.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.71.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.71.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.71.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.71.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.71.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.71.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.71.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.71.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.71.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.71.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.71.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.71.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.72.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.72.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.72.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.72.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.72.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.72.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.72.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.72.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.72.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.72.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.72.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.72.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.72.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.72.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.72.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.72.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.72.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.72.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.72.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.72.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.72.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.72.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.72.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.73.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.73.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.73.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.73.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.73.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.73.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.73.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.73.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.73.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.73.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.73.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.73.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.73.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.73.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.73.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.73.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.73.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.73.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.73.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.73.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.73.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.73.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.73.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.74.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.74.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.74.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.74.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.74.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.74.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.74.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.74.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.74.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.74.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.74.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.74.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.74.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.74.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.74.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.74.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.74.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.74.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.74.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.74.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.74.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.74.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.74.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.75.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.75.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.75.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.75.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.75.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.75.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.75.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.75.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.75.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.75.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.75.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.75.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.75.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.75.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.75.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.75.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.75.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.75.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.75.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.75.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.75.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.75.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.75.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.76.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.76.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.76.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.76.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.76.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.76.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.76.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.76.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.76.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.76.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.76.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.76.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.76.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.76.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.76.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.76.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.76.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.76.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.76.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.76.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.76.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.76.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.76.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.77.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.77.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.77.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.77.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.77.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.77.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.77.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.77.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.77.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.77.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.77.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.77.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.77.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.77.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.77.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.77.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.77.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.77.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.77.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.77.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.77.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.77.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.77.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.78.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.78.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.78.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.78.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.78.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.78.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.78.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.78.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.78.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.78.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.78.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.78.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.78.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.78.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.78.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.78.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.78.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.78.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.78.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.78.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.78.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.78.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.78.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.79.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.79.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.79.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.79.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.79.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.79.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.79.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.79.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.79.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.79.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.79.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.79.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.79.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.79.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.79.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.79.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.79.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.79.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.79.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.79.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.79.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.79.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.79.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.80.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.80.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.80.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.80.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.80.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.80.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.80.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.80.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.80.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.80.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.80.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.80.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.80.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.80.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.80.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.80.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.80.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.80.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.80.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.80.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.80.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.80.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.80.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.81.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.81.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.81.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.81.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.81.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.81.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.81.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.81.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.81.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.81.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.81.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.81.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.81.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.81.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.81.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.81.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.81.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.81.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.81.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.81.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.81.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.81.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.81.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.82.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.82.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.82.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.82.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.82.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.82.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.82.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.82.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.82.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.82.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.82.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.82.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.82.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.82.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.82.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.82.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.82.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.82.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.82.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.82.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.82.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.82.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.82.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.83.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.83.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.83.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.83.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.83.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.83.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.83.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.83.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.83.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.83.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.83.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.83.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.83.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.83.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.83.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.83.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.83.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.83.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.83.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.83.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.83.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.83.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.83.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.84.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.84.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.84.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.84.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.84.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.84.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.84.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.84.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.84.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.84.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.84.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.84.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.84.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.84.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.84.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.84.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.84.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.84.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.84.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.84.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.84.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.84.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.84.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.85.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.85.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.85.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.85.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.85.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.85.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.85.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.85.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.85.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.85.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.85.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.85.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.85.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.85.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.85.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.85.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.85.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.85.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.85.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.85.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.85.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.85.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.85.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.86.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.86.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.86.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.86.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.86.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.86.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.86.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.86.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.86.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.86.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.86.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.86.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.86.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.86.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.86.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.86.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.86.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.86.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.86.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.86.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.86.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.86.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.86.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.87.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.87.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.87.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.87.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.87.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.87.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.87.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.87.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.87.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.87.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.87.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.87.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.87.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.87.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.87.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.87.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.87.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.87.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.87.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.87.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.87.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.87.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.87.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.88.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.88.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.88.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.88.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.88.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.88.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.88.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.88.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.88.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.88.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.88.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.88.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.88.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.88.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.88.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.88.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.88.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.88.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.88.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.88.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.88.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.88.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.88.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.89.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.89.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.89.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.89.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.89.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.89.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.89.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.89.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.89.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.89.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.89.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.89.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.89.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.89.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.89.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.89.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.89.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.89.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.89.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.89.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.89.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.89.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.89.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.90.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.90.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.90.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.90.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.90.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.90.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.90.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.90.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.90.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.90.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.90.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.90.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.90.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.90.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.90.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.90.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.90.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.90.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.90.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.90.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.90.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.90.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.90.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.91.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.91.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.91.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.91.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.91.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.91.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.91.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.91.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.91.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.91.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.91.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.91.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.91.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.91.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.91.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.91.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.91.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.91.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.91.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.91.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.91.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.91.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.91.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.92.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.92.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.92.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.92.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.92.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.92.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.92.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.92.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.92.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.92.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.92.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.92.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.92.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.92.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.92.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.92.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.92.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.92.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.92.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.92.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.92.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.92.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.92.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.93.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.93.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.93.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.93.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.93.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.93.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.93.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.93.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.93.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.93.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.93.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.93.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.93.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.93.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.93.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.93.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.93.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.93.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.93.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.93.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.93.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.93.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.93.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.94.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.94.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.94.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.94.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.94.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.94.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.94.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.94.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.94.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.94.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.94.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.94.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.94.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.94.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.94.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.94.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.94.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.94.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.94.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.94.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.94.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.94.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.94.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.95.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.95.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.95.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.95.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.95.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.95.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.95.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.95.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.95.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.95.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.95.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.95.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.95.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.95.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.95.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.95.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.95.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.95.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.95.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.95.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.95.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.95.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.95.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.96.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.96.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.96.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.96.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.96.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.96.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.96.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.96.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.96.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.96.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.96.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.96.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.96.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.96.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.96.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.96.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.96.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.96.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.96.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.96.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.96.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.96.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.96.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.97.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.97.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.97.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.97.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.97.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.97.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.97.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.97.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.97.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.97.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.97.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.97.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.97.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.97.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.97.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.97.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.97.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.97.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.97.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.97.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.97.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.97.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.97.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.98.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.98.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.98.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.98.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.98.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.98.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.98.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.98.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.98.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.98.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.98.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.98.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.98.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.98.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.98.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.98.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.98.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.98.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.98.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.98.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.98.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.98.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.98.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.99.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.99.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.99.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.99.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.99.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.99.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.99.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.99.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.99.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.99.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.99.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.99.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.99.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.99.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.99.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.99.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.99.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.99.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.99.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.99.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.99.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.99.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.99.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.100.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.100.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.100.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.100.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.100.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.100.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.100.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.100.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.100.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.100.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.100.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.100.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.100.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.100.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.100.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.100.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.100.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.100.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.100.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.100.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.100.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.100.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.100.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.101.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.101.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.101.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.101.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.101.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.101.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.101.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.101.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.101.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.101.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.101.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.101.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.101.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.101.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.101.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.101.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.101.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.101.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.101.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.101.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.101.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.101.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.101.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.102.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.102.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.102.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.102.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.102.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.102.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.102.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.102.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.102.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.102.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.102.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.102.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.102.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.102.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.102.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.102.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.102.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.102.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.102.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.102.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.102.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.102.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.102.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.103.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.103.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.103.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.103.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.103.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.103.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.103.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.103.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.103.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.103.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.103.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.103.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.103.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.103.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.103.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.103.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.103.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.103.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.103.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.103.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.103.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.103.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.103.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.104.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.104.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.104.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.104.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.104.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.104.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.104.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.104.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.104.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.104.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.104.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.104.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.104.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.104.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.104.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.104.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.104.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.104.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.104.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.104.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.104.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.104.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.104.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.105.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.105.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.105.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.105.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.105.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.105.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.105.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.105.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.105.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.105.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.105.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.105.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.105.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.105.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.105.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.105.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.105.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.105.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.105.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.105.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.105.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.105.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.105.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.106.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.106.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.106.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.106.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.106.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.106.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.106.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.106.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.106.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.106.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.106.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.106.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.106.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.106.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.106.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.106.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.106.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.106.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.106.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.106.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.106.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.106.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.106.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.107.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.107.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.107.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.107.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.107.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.107.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.107.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.107.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.107.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.107.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.107.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.107.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.107.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.107.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.107.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.107.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.107.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.107.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.107.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.107.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.107.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.107.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.107.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.108.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.108.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.108.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.108.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.108.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.108.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.108.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.108.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.108.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.108.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.108.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.108.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.108.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.108.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.108.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.108.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.108.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.108.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.108.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.108.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.108.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.108.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.108.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.109.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.109.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.109.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.109.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.109.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.109.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.109.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.109.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.109.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.109.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.109.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.109.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.109.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.109.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.109.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.109.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.109.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.109.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.109.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.109.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.109.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.109.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.109.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.110.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.110.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.110.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.110.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.110.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.110.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.110.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.110.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.110.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.110.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.110.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.110.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.110.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.110.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.110.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.110.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.110.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.110.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.110.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.110.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.110.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.110.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.110.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.111.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.111.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.111.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.111.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.111.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.111.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.111.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.111.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.111.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.111.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.111.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.111.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.111.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.111.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.111.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.111.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.111.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.111.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.111.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.111.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.111.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.111.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.111.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.112.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.112.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.112.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.112.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.112.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.112.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.112.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.112.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.112.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.112.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.112.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.112.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.112.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.112.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.112.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.112.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.112.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.112.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.112.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.112.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.112.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.112.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.112.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.113.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.113.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.113.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.113.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.113.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.113.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.113.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.113.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.113.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.113.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.113.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.113.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.113.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.113.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.113.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.113.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.113.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.113.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.113.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.113.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.113.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.113.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.113.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.114.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.114.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.114.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.114.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.114.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.114.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.114.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.114.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.114.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.114.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.114.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.114.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.114.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.114.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.114.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.114.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.114.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.114.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.114.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.114.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.114.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.114.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.114.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.115.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.115.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.115.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.115.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.115.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.115.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.115.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.115.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.115.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.115.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.115.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.115.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.115.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.115.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.115.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.115.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.115.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.115.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.115.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.115.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.115.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.115.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.115.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.116.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.116.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.116.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.116.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.116.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.116.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.116.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.116.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.116.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.116.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.116.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.116.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.116.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.116.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.116.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.116.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.116.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.116.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.116.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.116.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.116.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.116.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.116.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.117.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.117.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.117.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.117.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.117.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.117.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.117.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.117.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.117.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.117.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.117.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.117.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.117.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.117.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.117.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.117.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.117.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.117.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.117.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.117.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.117.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.117.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.117.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.118.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.118.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.118.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.118.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.118.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.118.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.118.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.118.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.118.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.118.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.118.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.118.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.118.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.118.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.118.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.118.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.118.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.118.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.118.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.118.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.118.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.118.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.118.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.119.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.119.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.119.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.119.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.119.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.119.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.119.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.119.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.119.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.119.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.119.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.119.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.119.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.119.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.119.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.119.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.119.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.119.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.119.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.119.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.119.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.119.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.119.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.120.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.120.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.120.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.120.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.120.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.120.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.120.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.120.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.120.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.120.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.120.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.120.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.120.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.120.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.120.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.120.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.120.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.120.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.120.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.120.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.120.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.120.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.120.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.121.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.121.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.121.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.121.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.121.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.121.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.121.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.121.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.121.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.121.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.121.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.121.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.121.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.121.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.121.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.121.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.121.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.121.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.121.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.121.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.121.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.121.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.121.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.122.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.122.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.122.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.122.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.122.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.122.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.122.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.122.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.122.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.122.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.122.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.122.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.122.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.122.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.122.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.122.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.122.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.122.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.122.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.122.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.122.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.122.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.122.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.123.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.123.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.123.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.123.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.123.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.123.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.123.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.123.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.123.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.123.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.123.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.123.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.123.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.123.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.123.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.123.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.123.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.123.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.123.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.123.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.123.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.123.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.123.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.124.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.124.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.124.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.124.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.124.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.124.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.124.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.124.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.124.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.124.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.124.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.124.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.124.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.124.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.124.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.124.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.124.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.124.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.124.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.124.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.124.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.124.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.124.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 }, { "quantization": "fp8", "decoder_type": "llama", "input_layernorm": { "weight": "model.layers.125.input_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp_layernorm": null, "attention": { "qkv": { "weight": "model.layers.125.self_attn.qkv.weight", "bias": null, "activation_scaling_factor": "model.layers.125.self_attn.qkv.input_quant_scale", "weights_scaling_factor": "model.layers.125.self_attn.qkv.weight_quant_scale", "output_scaling_factor": "model.layers.125.self_attn.qkv.output_quant_scale", "awq_block_size": 0 }, "kv_cache_scaling_factor": "model.layers.125.kv_cache_scaling_factor", "kv_cache_dtype": "FP8", "dense": { "weight": "model.layers.125.self_attn.o_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.125.self_attn.o_proj.input_quant_scale", "weight_scaling_factor": "model.layers.125.self_attn.o_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.125.self_attn.o_proj.output_quant_scale" } }, "post_layernorm": { "weight": "model.layers.125.post_attention_layernorm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "mlp": { "fc": { "weight": "model.layers.125.mlp.gate_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.125.mlp.gate_proj.input_quant_scale", "weight_scaling_factor": "model.layers.125.mlp.gate_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.125.mlp.gate_proj.output_quant_scale" }, "gate": { "weight": "model.layers.125.mlp.up_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.125.mlp.up_proj.input_quant_scale", "weight_scaling_factor": "model.layers.125.mlp.up_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.125.mlp.up_proj.output_quant_scale" }, "proj": { "weight": "model.layers.125.mlp.down_proj.weight", "bias": null, "activation_scaling_factor": "model.layers.125.mlp.down_proj.input_quant_scale", "weight_scaling_factor": "model.layers.125.mlp.down_proj.weight_quant_scale", "awq_block_size": 0, "output_scaling_factor": "model.layers.125.mlp.down_proj.output_quant_scale" }, "hidden_act": "silu" }, "num_attention_heads": 128, "attention_head_size": 128, "num_kv_heads": 8, "max_position_embeddings": 131072, "rotary_pct": 0, "parallel_attention": false, "apply_residual_connection_post_layernorm": false, "use_cache": true, "rope_ratio": 1.0, "seq_length": 0 } ], "final_layernorm": null, "ln_f": { "weight": "model.norm.weight", "bias": null, "layernorm_type": "rms", "eps": 1e-05 }, "lm_head": { "weight": "lm_head.weight", "bias": null, "activation_scaling_factor": null, "weight_scaling_factor": null, "awq_block_size": 0, "output_scaling_factor": null }, "share_embedding_table": false }